comparative evaluation of interactive segmentation algorithms … · 2019. 6. 13. · comparative...

International Journal of Computer Information Systems and Industrial Management Applications.

ISSN 2150-7988 Volume 11 (2019) pp. 142-150

© MIR Labs, www.mirlabs.net/ijcisim/index.html

MIR Labs, USA

Received: 17 Feb, 2019; Accepted: 15 May, 2019; Published: 13 June, 2019

Comparative Evaluation of Interactive Segmentation

Algorithms Using One Unified User Interactive Type

Soo See Chai1, Kok Luong Goh2, Hui Hui Wang3 and Yin Chai Wang4

1,3,4 Faculty of Computer Science & Information Technology,

University of Malaysia Sarawak (UNIMAS),

Kota Samarahan, 94300, Malaysia [email protected], [email protected], [email protected]

2 School of Innovative Technology,

International College of Advanced Technology Sarawak (i-CATS),

Malaysia

[email protected]

Abstract: In interactive segmentation, user inputs are required

to produce cues for the algorithms to extract the object of interest.

Different input types were recommended by the researchers in

their developed algorithms. The most common input types are

points, strokes and bounding box. Different evaluation

parameters were used in the researches in this field for

comparison. Our previous work shows that, for non-complex

image, segmentation result will not be affected by the user input

type used. Complex images are defined as images whereby the

colors of the objects of interest and the background are similar

and vice-versa. In some of the complex images, parts of the color

of the objects of interest are present in the background. This

paper extends our previous work by using the proposed unified

input types, which consists of a bounding box to locate the object

of interest range and a stroke for the foreground, on three

interactive segmentation algorithms for non-complex and

complex image. Three different evaluation measures are

computed to compare the segmentation quality: Variation of

Information (VI), Global Consistency Error (GCE) and Jaccard

index (JI). From the experiment results, it is noticed that, all

three algorithms perform well for non-complex images but could

not perform as good for complex images.

Keywords: interactive segmentation, complex, non-complex, user

input, bounding box, stroke.

I. Introduction

Image segmentation algorithms help human to extract object

of interest from images for further processing. Generally,

image segmentation will partition an image into certain

number of regions which have certain coherent features like

texture and colors. These coherent features would be grouped

into meaningful pieces for better perceiving [1]. From the

technical perspective, image segmentation can be divided into:

fully-automated, semi-automated or interactive, and manual

segmentation [2]. Fully automated segmentation, as the name

implies, does not require user intervention. Semi-automated

or interactive segmentation, where user automation is at

medium, requires user to initialize the algorithms to

appropriately mark the object of interest. Some of these

algorithms will require user to provide feedback to the

algorithms in order to improve the segmentation results. In

manual segmentation, the object of interest is delineated by

hand. In most of the practical applications of image

segmentation, large number of images are needed to be

handled by human. Therefore, human intervention in the

segmentation process should be as minimal as possible. This

makes automated image segmentation more appealing [3].

However, automated segmentation still exhibits certain

constraints and cannot produce satisfactory results due to the

complexity of the images, especially using natural images

[4-7]. To overcome this, semi-automated or interactive

segmentation use human operator to provide cue to the

segmentation algorithms.

This paper is organized as follows: Section II. will present a

brief introduction on the general interactive segmentation and

the three different algorithms used. The purpose of this paper

is also explained in this section. Section III. presents the

experiment settings and the results obtained. Conclusion are

presented in Section V.

II. Interactive Segmentation

In interactive segmentation, the user intention is incorporated

in the user interaction [8]. The user will provide intuitive

interaction which serves as high level or prior information to

the interactive segmentation algorithms to extract the object of

interest [9-14]. A general functional view of an interactive

segmentation system is depicted in Figure 1. The system

consists of User Input Module (Step 1), Computation Module

(Step 2) and Output Display Module (Step 3) [15]. In Step 1,

the module will receive user input and the intention of the user

will be recognized at this stage. The main part of the whole

system is the Computation Module, i.e Step 2, whereby the

segmentation algorithms will run according to the user input to

generate the segmentation results. The segmentation results

will be displayed in Step 3. These steps are iterative whereby

Comparative Evaluation of Interactive Segmentation Algorithms Using One Unified User Interactive Type 143

the user can input additional input after Step 3 and the system

will go back to Step 1 until

Figure 1. General functional view of an interactive

segmentation system whereby a user can produce interaction

to the system until a satisfactory result [15].

Figure 2. Input types used in interactive segmentation: a:

bounding box. b: bounding boxes for foreground object. c:

seed points for the background and foreground of the image. d:

placing the skeleton on the object of interest. e: background

and foreground strokes on the image. f: stroke on the contour

of the object. g: bounding box with strokes on the object and

background, and h: seed point on the contour of the object.

the user satisfies with the output produced. Thus, this is a

human-machine collaborative method whereby, the machine

will need to understand the user intention based on the human

interaction and the human’s input will affect the machine

behavior in solving the problem [16].

There are several user interactive or input types used in the

current researches to offer the information about the

background and foreground regions. The most commonly

used input type is placing strokes in the foreground and

background in the image [4, 6, 7, 9, 13, 17-23]. Besides

strokes, rectangle or bounding box is also used to locate the

target object range [11, 24-28]. In addition, seed points are

also used to be placed on foreground and background on the

image [29-31] or on the contour of the object of interest [32,

33]. Mixed input types were also applied in some of the

research. For example, [34] combined seed points with strokes

on the background and foreground of the image to extract the

object of the interest. Figure 2 shows the use of different input

types used in interactive segmentation.

Figure 3. The overall framework of [38].

In our previous work [35], we had compared four

commonly used user input types, i.e, seed point, foreground

and background strokes, foreground stroke with background

bounding box, and bounding box as foreground and

background using the nonparametric higher-order learning

algorithm by [36] for the simple or non-complex and complex

images from the Berkeley image database [37]. Our previous

findings show that, the use of bounding box to locate the object

of interest range can improve the segmentation results for

complex image. In addition, we also found that, the location

and length of strokes or seed points used have an impact on the

segmentation accuracy. In this work, we will extend the

previous work by comparing and evaluating the use of

bounding box as input type for three different interactive

segmentation algorithms: robust interactive image

segmentation via graph-based manifold ranking [38],

f

f.

a.

b.

c.

e.

g.

Chai et al.

144

nonparametric higher-order learning for interactive

segmentation [36] and interactive image segmentation by

maximal similarity based region merging [39]. Quantitative

comparison evaluation on the segmentation results will be

done with three evaluation parameters: Variation of

Information (VI), Global Consistency Error (GCE) and

Jaccard index (JI) [40]. Between two segmentation, the GCE

is the error measure, JI measures the similarity and VI

measures the distance.

a. User scribbles f. Output with = 0.9665

b. User scribbles g. Output with = 0.9644

c. User scribbles h. Output with = 0.9491

d. User scribbles i. Output with = 0.9682

e. User scribbles j. Output with = 0.9627

Figure 4. Segmentation results obtained using [38]: a. to e.

multiple user scribbles for foreground and background labels,

and f. to j. the corresponding result obtain.

a. f

b. g.

c. h.

d. i.

e. j.

Figure 5. Segmentation results obtained using algorithm [36]:

a. to e. user scribbles used, f. to j. the corresponding results.

A. Robust Interactive Segmentation via Graph-based

Manifold Ranking [38]

In this work, using graph-based semi-supervised learning

theory and superpixel, an interactive segmentation system was

developed. At the beginning of the process, the image is

over-segmented into small homogeneous regions using

superpixels. To cater the user intention, scribbles are entered

as the foreground and background labels. There is no limit on

the amount of foreground and background scribbles that may

be input by the user. These foreground and background labels


information will be integrated into the superpixels. Next, using

the proposed labels driven and locally adaptive kernel

parameter, the k-regular sparse graph is approximated to form

the affinity graph matrix. By calculating and integrating the

ranking scores, the final segmentation result is generated to

form the foreground and background indicator vectors from

the user scribbles. The overall framework is shown in Figure 3.

Figure 4 shows some of the segmentation results the authors

obtained. In their work, the normalized overlap, β0, is used to

measure the similarity between the segmentation results and

the preset ground truth. However, the average β0 obtained

from all the tested images was not reported. In addition, the

authors reported that, the proposed algorithm managed to

segment the input images in less than 2 seconds.

The key contributions of this work are:

• A novel framework that combines the graph-based

semi-supervised learning theory with region-based

models to efficiently segment out the desired object,

• A novel graph construction strategy which models the

graph as an approximate k-regular sparse graph that

integrate spatial relationships and user provided

scribbles, and

• A new graph edge weights computing strategy that forms

the weights using a locally adaptive kernel width

parameter.

B. Nonparametric higher-order learning for interactive

segmentation [36]

This generative interactive model was based on multi labels.

An algorithm was proposed to estimate the pixels likelihood

for each label. Using the mean shift unsupervised learning

algorithm [41], a new higher-order cost function of pixel

likelihoods to partly enforce the label consistency inside the

regions generated was designed. The algorithm considers the

relationships between the pixels and the corresponding regions

in a multi-layer graph. To estimate the pixels likelihoods, the

higher-order cues of the representative region likelihoods are

used. The non-parametric learning technique is introduced to

recursively estimate the higher-order cues from the resulting

likelihoods of pixels included in each region. This will

consider the connections between the regions that aid in the

propagation of the local grouping cues across the image areas.

Figure 5 shows the scribbles used and the corresponding

segmentation results for some of the images tested by the

authors. The authors reported that the algorithm resulted in

4.34% of error rate.

The key contributions of this algorithm are:

• The pixel likelihood for each label is estimated,

• The representative region likelihoods are defined as

higher-order cues to estimate the pixel likelihoods, and

• Using the resulting likelihoods of pixels included in each

region, a nonparametric learning technique was

addressed to recursively estimate the higher order cues.

For the pixel and region properties, only color values were

used.

C. Interactive image segmentation by maximal

similarity-based region merging [39]

For this method, strokes are used not only as the markers to

indicate the location, but also to label the object and

background. For the different regions, the similarity between

these regions are calculated. Utilizing the assistance of the

markers input by the user, the similar regions are merged based

on the proposed maximal similarity rule. This proposed

maximal similarity-based region merging does not require a

preset threshold value and is adaptive to image content. The

method manages to merge and label the non-marker

background automatically.

a. f.

b. g.

c. h.

d. i.

e. j.

Figure 6. Segmentation results obtained using algorithm [39]:

a. to e. show the user input, f.to j. show the segmentation

outputs obtain.

Chai et al.

146

In addition, the method also successfully determined the

non-marker object regions, and these are avoided from being

merged with the background. The object contours are next

extracted from the background after all the non-marker regions

are labeled. The authors claimed that this method is the first to

use the markers entered by the user to guide the region

merging for object contour extraction. Some of the results

obtained using this method are shown in Figure 6.

The authors manually labelled the desired object in the

images and used this as the ground truth. The evaluation

parameters used are: true positive rate and false positive rate.

True positive rate is defined as the number of correctly

classified object pixels over the total number of object pixels

in the ground truth. The ratio of the number of background

pixels which are wrongly classified as object pixels over the

total number of background pixels in the ground truth is the

false positive rate.

a.

b.

Figure 7. Failure example of algorithm [39]: a. Part of the

object and background have very similar features and many

scribbles were use, b. segmentation result obtained.

The ultimate aim is to achieve lower false positive rate

and at the same time, higher true positive rate. The true

positive rate and false positive rate obtained for the images in

Figure 6 are: bird (94.64% and 0.29%), starfish (90.25% and

0.26%), flower (97.59% and 1.08%), Mona Lisa (98.85%,

0.71%) and tiger (91.70%, 0.75%).

The author had noted that, since the proposed algorithm

is based on some initial segmentations, such as mean-shift or

superpixel, therefore, if the initial segmentation does not

provide a good basis of region merging, the proposed method

will fail.

In terms of user input, the experiments conducted by the

authors showed that, the user input had to cover the main

feature regions in odder for the object of interest to be

correctly extracted. In addition, the authors had also

conducted experiments on the different number of user input.

It was concluded that, the images with more user inputs will

produce better segmentation results. The algorithm was

noticed to produce not so encouraging results when shadow,

low-contrast edges and ambiguous areas occur. In an

experiment using an image where parts of the object regions

are very similar to the background, as shown in Figure 7, the

segmentation result obtained was poor although many

markers were used.

Figure 8. The bounding box and foreground stroke used as the

input image for the three different interactive segmentation

algorithms. Image a, b and c are non-complex images used

while, d, e and f images are in the category of complex images.

a. d.

b. e.

c. f.


III. Experiment Settings and Results

A. Observations

From the previous section, it can be seen that, the three

algorithms use strokes to indicate the foreground and

background in the image. Multiple inputs are required from the

user to accurately provide the cue for the algorithms for

segmentation purpose. In addition, there is no common

evaluation features used. Therefore, in this study, the

experiments settings will: 1) compare the three algorithms

based on the unified single bounding box and one stroke for

the foreground in a set of images that are divided into complex

and non-complex categories, and 2) use the same evaluation

parameters to compare and evaluate the segmentation results.

GCE=0.02

VI=0.98

JI=0.95

GCE=0.01

VI=0.99

JI=0.98

GCE=0.01

VI=0.99

JI=0.98

GCE=0.12

VI=0.87

JI=0.83

GCE=0.02

VI=0.98

JI=0.97

GCE=0.04

VI=0.96

JI=0.95

GCE=0.11

VI=0.88

JI=0.86

GCE=0.02

VI=0.97

JI=0.97

GCE=0.04

VI=0.96

JI=0.95

a. b. c.

Figure 9. Qualitative and quantitative segmentation results

obtain using non-complex images for: a. Robust Interactive

Segmentation via Graph-based Manifold Ranking [38], b.

Nonparametric higher-order learning for interactive

segmentation [36], and c. Interactive image segmentation by

maximal similarity-based region merging [39].

B. Dataset

The Berkeley image database [37] consists of around 12,000

hand-labeled segmentations of 1,000 Corel dataset from 30

human subjects. This database was available publicly for

research on image segmentation and boundary. Using the

images from the Berkeley image database [37], the images are

divided into non-complex and complex images. Non-complex

images are images whereby the color of the objects of interest

and background are not similar while for complex images, the

objects of interest and the background are having akin color.

In some of the complex images, parts of the color of the object

of interest are present in the background.

In the first stage, the user input which includes the bounding

box to locate the range of the object of interest and a stroke for

the foreground will be drawn on the selected images. Next,

these user scribbled images are input into the three different

interactive segmentation algorithms. The three evaluation

parameters: Variation of Information (VI), Global

Consistency Error (GCE) and Jaccard index (JI) are calculated

for the segmentation results.

GCE=0.07

VI=0.88

JI=0.50

GCE=0.04

VI=0.91

JI=0.37

GCE=0.05

VI=0.94

JI=0.62

GCE=0.13

VI=0.82

JI=0.62

GCE=0.08

VI=0.81

JI=0.36

GCE=0.07

VI=0.92

JI=0.77

GCE=0.10

VI=0.77

JI=0.34

GCE=0.04

VI=0.93

JI=0.55

GCE=0.03

VI=0.95

JI=0.70

a. b. c.

Figure 10. Qualitative and quantitative segmentation results

obtain using complex images for: a. Robust Interactive

Segmentation via Graph-based Manifold Ranking [38], b.

Nonparametric higher-order learning for interactive

segmentation [36], and c. Interactive image segmentation by

maximal similarity based region merging [39].

Chai et al.

148

C. Results and Discussions

For comparison purpose, three non-complex images and

three complex images were used. Figure 8 shows the images

in these two categories. In this figure, image a. to c. are

non-complex images whereby the background and object of

interest could be differentiated easily. Images d. to f. in this

figure are complex images. In these complex images, it could

be seen that, parts of the objects and the background have very

similar color features. The unified bounding box and

foreground stroke from the user input is also shown in this

figure. Fair comparison is done in the experiments as the

length and location of the strokes as well as the size of the

bounding box use for the three algorithms are the same. It is

worth mentioning that, the over-segmentation technique used

in each of the algorithm is remained the same, i.e. there is no

change on the original algorithm in terms of

over-segmentation technique used.

1) Non-complex Images

For the non-complex images, the segmentation results

obtained using the three interactive segmentation algorithms

are shown in Figure 9. Using the three evaluation parameters,

it can be concluded that, the three interactive segmentation

algorithms can extract the object of interest with satisfactory

results, with minimum JI=0.83 and GCE=0.12. The best

segmentation results obtained was JI=0.98 and GCE=0.01.

The Nonparametric higher-order learning for interactive

segmentation (NHL) [36], and Interactive image segmentation

by maximal similarity based region merging (MSRM) [39]

perform better as comparing to the Robust Interactive

Segmentation via Graph-based Manifold Ranking (RGMR)

[38] with NHL and MSRM achieve more than or equal to 0.95

for JI and GCE less than 0.05. RGMR on the other hand, only

managed to achieve the best result using the bush image with

JI=0.95 and GCE=0.02. For the other two non-complex

images, RGMR only manages to achieve average JI of around

0.85 and GCE of around 0.12. With visual inspection, it could

be seen that, the objects extracted mostly do not include the

background of the image for the three algorithms. However,

the detail of the object, for example, the image of the vase, is

partly missed.

2) Complex Images

Using the complex images, as shown in Figure 10, all three

algorithms could not perform as good as using the

non-complex images. The best interactive segmentation

algorithm among these three algorithms is the MSRM which

could still produce segmentation result of JI more than 0.60

and GCE of less than 0.10 for all the three complex images

used. RGMR produces the highest GCE for all three complex

images as comparing to the other two algorithms. In terms of

GCE, the ranking of the algorithms from low to high values is:

MSRM, NHL and RGMR. However, if using JI, the ranking is:

NHL, RGMR and MSRM. With visual inspection, it could be

seen that, the segmentation includes the background of the

object, which means that, the three algorithms fail to

differentiate between the object of interest and the background

for complex images.

IV. Conclusion

This paper uses the unified bounding box to locate the ranhe of

the object of interest and a stroke in the foreground as user

input for the three interactive segmentations: RGMR, NHL

and MSRM. The images used are divided into non-complex

and complex images. For the images used, the unified input

type suggested in this paper, which is a bounding box to locate

the object of interest and a stroke placed on the foreground of

the images, were drawn. In another words, the location and

size/length of the bounding box and stroke are the same for all

the test images. In addition, to better compare and evaluate the

results, three evaluation parameters are used for all the three

interactive algorithms, i.e. GCE, VI and JI. It was found that,

all three interactive segmentation algorithms perform well for

non-complex images with JI values of more than 0.80. MSRM

and NHL outperform RGMR for all the three non-complex

images. However, for complex images, only MSRM can

achieve JI values of more than 0.60 and GCE values of less

than 0.10. RGMR performs the worst among these three

algorithms for complex images with highest GCE value

obtained only at 0.13 and lowest JI value at 0.34. It can be

concluded that, the MSRM algorithm can produce more

consistent segmentation results as comparing to the NHL and

RGMR algorithms, although for complex images, the results

obtained using MSRM deteriorates as comparing to extracting

objects of interest in non-complex images. In addition, for

MSRM, it can also be generalized that, the influences of the

user input type on the segmentation results is not as huge as the

NHL and RGMR algorithms for complex images. Visual

inspection of the segmentation results reveal that, for complex

images, the three interactive segmentation algorithms fail to

differentiate between the object of interest with the

background and this result in background being included in the

final segmentation results.

The proposed unified bounding box and single stroke used

in this research suggest a method to extract objects of interest

with less required input. However, from the experiments in

this study, it can be concluded that, all the three interactive

segmentations could not perform well in extracting object of

interest for complex images, but the extraction of object of

interest is good when using non-complex images. In the future,

we would experiment with more complex images and more

interactive segmentation algorithms to verify the results

obtained.

Acknowledgment

The authors would like to thank the Faculty of Computer

Science and Information Technology (FCSIT), University

Malaysia Sarawak (UNIMAS) for providing the facilities

needed.

References

[1] Forsyth, D.A. and J. Ponce, Computer vision: a modern

approach2002: Prentice Hall Professional Technical

Reference.


[2] Heckel, F., et al., On the evaluation of segmentation

editing tools. Journal of Medical Imaging, 2014. 1(3): p.

034005.

[3] Peng, B., L. Zhang, and D. Zhang, Automatic image

segmentation by dynamic region merging. IEEE

Transactions on image processing, 2011. 20(12): p.

3592-3605.

[4] Wang, F., X. Wang, and T. Li. Efficient label propagation

for interactive image segmentation. in Machine Learning

and Applications, 2007. ICMLA 2007. Sixth International

Conference on. 2007. IEEE.

[5] Jia, Y. and C. Zhang. Learning distance metric for

semi-supervised image segmentation. in Image

Processing, 2008. ICIP 2008. 15th IEEE International


[6] Jiang, X., A. Große, and K. Rothaus, Interactive

segmentation of non-star-shaped contours by dynamic

programming. Pattern Recognition, 2011. 44(9): p.

2008-2016.

[7] Nguyen, T.-V., et al. Efficient image segmentation

incorporating photometric and geometric information. in

Proceedings of the International MultiConference of

Engineers and Computer Scientists. 2011.

[8] Yang, W., et al., User-friendly interactive image

segmentation through unified combinatorial user inputs.

IEEE Transactions on image processing, 2010. 19(9): p.

2470-2479.

[9] Xu, J., X. Chen, and X. Huang. Interactive image

segmentation by semi-supervised learning ensemble. in

Knowledge Acquisition and Modeling, 2008. KAM'08.

International Symposium on. 2008. IEEE.

[10] McGuinness, K. and N.E. O’connor, A comparative

evaluation of interactive segmentation algorithms.

Pattern Recognition, 2010. 43(2): p. 434-444.

[11] Vicente, S., V. Kolmogorov, and C. Rother. Joint

optimization of segmentation and appearance models. in

Computer Vision, 2009 IEEE 12th International


[12] Santner, J., T. Pock, and H. Bischof. Interactive

multi-label segmentation. in Asian Conference on

Computer Vision. 2010. Springer.

[13] Gulshan, V., et al. Geodesic star convexity for interactive

image segmentation. in Computer Vision and Pattern

Recognition (CVPR), 2010 IEEE Conference on. 2010.

IEEE.

[14] Nickisch, H., et al. Learning an interactive segmentation

system. in Proceedings of the Seventh Indian Conference

on Computer Vision, Graphics and Image Processing.

2010. ACM.

[15] Malmberg, F., Graph-based methods for interactive

image segmentation, 2011, Acta Universitatis

Upsaliensis.

[16] He, J., C.-S. Kim, and C.-C.J. Kuo, Interactive

Segmentation Techniques: Algorithms and Performance

Evaluation2013: Springer Science & Business Media.

[17] Vezhnevets, V. and V. Konouchine. GrowCut:

Interactive multi-label ND image segmentation by

cellular automata. in proc. of Graphicon. 2005. Citeseer.

[18] Mu, Y., B. Zhou, and S. Yan, Information-theoretic

analysis of input strokes in visual object cutout. IEEE

Transactions on Multimedia, 2010. 12(8): p. 843-852.

[19] Wu, X. and Y. Wang. Interactive foreground/background

segmentation based on graph cut. in Image and Signal

Processing, 2008. CISP'08. Congress on. 2008. IEEE.

[20] Liu, M., S. Chen, and J. Liu. Precise object cutout from

images. in Proceedings of the 16th ACM international

conference on Multimedia. 2008. ACM.

[21] Kim, T.H., K.M. Lee, and S.U. Lee. Generative image

segmentation using random walks with restart. in

European conference on computer vision. 2008.

Springer.

[22] Liu, L., et al., A variational model and graph cuts

optimization for interactive foreground extraction. Signal

Processing, 2011. 91(5): p. 1210-1215.

[23] Xiang, S., et al., Interactive image segmentation with

multiple linear reconstructions in windows. IEEE

Transactions on Multimedia, 2011. 13(2): p. 342-352.

[24] Lempitsky, V., et al. Image segmentation with a bounding

box prior. in Computer Vision, 2009 IEEE 12th

International Conference on. 2009. IEEE.

[25] Rother, C., V. Kolmogorov, and A. Blake. Grabcut:

Interactive foreground extraction using iterated graph

cuts. in ACM transactions on graphics (TOG). 2004.

ACM.

[26] Liu, D., et al. Fast interactive image segmentation by

discriminative clustering. in Proceedings of the 2010

ACM multimedia workshop on Mobile cloud media

computing. 2010. ACM.

[27] Lee, S.H., H.I. Koo, and N.I. Cho, Image segmentation

algorithms based on the machine learning of features.

Pattern Recognition Letters, 2010. 31(14): p. 2325-2336.

[28] Friedland, G., K. Jantz, and R. Rojas. Siox: Simple

interactive object extraction in still images. in

Multimedia, Seventh IEEE International Symposium on.

2005. IEEE.

[29] Mičušík, B. and A. Hanbury. Steerable semi-automatic

segmentation of textured images. in Scandinavian

Conference on Image Analysis. 2005. Springer.

[30] Arbeláez, P. and L. Cohen. Constrained image

segmentation from hierarchical boundaries. in Computer

Vision and Pattern Recognition, 2008. CVPR 2008. IEEE


[31] Demjen, E. and J. Marek. Interactive algorithm for

accurate segmentation of fiber-like objects. in Image and

Signal Processing and Analysis, 2009. ISPA 2009.

Proceedings of 6th International Symposium on. 2009.

IEEE.

[32] Kang, H.W., G-wire: A livewire segmentation algorithm

based on a generalized graph formulation. Pattern

Recognition Letters, 2005. 26(13): p. 2042-2051.

[33] Zadicario, E., et al. Boundary snapping for robust image

cutouts. in Computer Vision and Pattern Recognition,

2008. CVPR 2008. IEEE Conference on. 2008. IEEE.

[34] de Miranda, P.A., A.X. Falcão, and J.K. Udupa,

Synergistic arc-weight estimation for interactive image

segmentation using graphs. Computer Vision and Image

Understanding, 2010. 114(1): p. 85-99.

[35] Chai, S.S. and K.L. Goh, Evaluation of Different User

Input Types in Interactive Segmentation. Journal of

Telecommunication, Electronic and Computer

Engineering (JTEC), 2017. 9(2-10): p. 91-99.

[36] Kim, T.H., K.M. Lee, and S.U. Lee. Nonparametric

higher-order learning for interactive segmentation. in

Chai et al.

150

Computer Vision and Pattern Recognition (CVPR), 2010

IEEE Conference on. 2010. IEEE.

[37] Martin, D., et al. A database of human segmented natural

images and its application to evaluating segmentation

algorithms and measuring ecological statistics. in

Computer Vision, 2001. ICCV 2001. Proceedings. Eighth

IEEE International Conference on. 2001. IEEE.

[38] Li, H., W. Wu, and E. Wu, Robust interactive image

segmentation via graph-based manifold ranking.

Computational Visual Media, 2015. 1(3): p. 183-195.

[39] Ning, J., et al., Interactive image segmentation by

maximal similarity based region merging. Pattern

Recognition, 2010. 43(2): p. 445-456.

[40] Sharma, M. and V. Chouhan, Objective evaluation

parameters of image segmentation algorithms.

International Journal of Engineering and Advanced

Technology (IJEAT), 2012. 2(2): p. 2249-8958.

[41] Comaniciu, D. and P. Meer, Mean shift: A robust

approach toward feature space analysis. IEEE

Transactions on pattern analysis and machine intelligence,

2002. 24(5): p. 603-619.

Author Biographies

Soo See Chai is presently working as a senior

lecturer at Department of Computing and

Software Engineering at Faculty of Computer

Science and Information Technology (FCSIT)

in University Malaysia Sarawak (UNIMAS).

Her research areas are GIS, Remote Sensing,

AI and Image Processing.

Kok Luong Goh obtained his Master of

Science in Information Technology from the

FCSIT, UNIMAS. He is currently working

as a lecturer in International College of

Advaned Technology Sarawak (ICATS),

Kuching.

Wang Hui Hui is currently a Senior Lecturer

at the Faculty of Computer Science and

Information Technology, University

Malaysia Sarawak (UNIMAS). She received

her Ph.D in Computer Science from the

Universiti Teknologi Malaysia. Her main

research interest is in image processing.

Wang Yin Chai is presently working as

Professor at Faculty of Computer Science

and Information Technology, UNIMAS since

2007. He had more than 24 years of

experience in research and development. The

areas of interests include AI, Image

processing and GIS Analysis. She had

published more than 160 publications and

leaded consultation projects amounting more

than 20 million.

comparative evaluation of interactive segmentation algorithms … · 2019. 6. 13. · comparative...

Documents